Amendo

Conversation with Amendo - net-new job
Conversation with Amendo - batch derivative job
Request confirmation sent by Amendo
Comments left by Amendo in processed form
GCP deployed services
GCP utilised models

Inspiration

Rent a car, rent a flat, book a vacation - and picture it showing up not quite matching what was promised: wrong spec, a missing extra, a condition nothing like the listing. You have to argue your case and fill a form.

The same cycle from the other side can be observed from the other side: someone fills out a form, someone else checks it against a checklist, and if anything's off, it bounces back with vague feedback and no explanation.

Multiply that by every insurance claim, fleet return, and inspection report in any office, and it's hours of unpaid attention nobody wants to give. We wanted to see if we could take that whole loop we experienced and put it behind nothing more than an email.

What it does

Send Amendo an email: your requirements as plain text, whatever paperwork or photos you've got as attachments. It writes back twice - first to confirm what it understood, then with the finished document. If you sent a filled-in form, it comes back annotated with what's fine and what isn't, untouched otherwise. If you only had photos and notes, it drafts the form for you, evidence attached. You can send both at once or in a batch.

How we built it (in two days)

Three services: email service polling the inbox and handling delivery, an editor service applying surgical changes on forms and a non-blocking computer vision service for image understanding

We were strict about keeping those separate - it meant we could test almost the whole system without ever calling an LLM, and grade the parts that do call one against a real submission instead of trusting our own gut. We wrote 156 evals against one real, human-photographed vehicle return.

One condensed day to get the pipeline working end to end, one day to get it standing up to a live demo - real inbox, real cloud infra.

We are especially proud of agent tooling which does surgical text edits (instead of regenerating) and parallel service for image understanding - these two really decrease the latency.

Challenges we ran into

The second day was almost entirely bugs we didn't expect mostly with regard to image understanding service. We had to learn how to spin up terraform CICD with Google Cloud. We somewhat bypassed the problem of creating our own email service by attaching the app to a preexisting one :)

What we learned

Write the checker at the same time as the feature (or before as in TDD).
Nothing you build is actually demo-ready until you've passed some messy input and watched what it fall over.